From life tables to language models

Guy Orcutt proposed modelling the economy as many individual people, households and firms, each making decisions, instead of only as aggregates.

The Simulmatics Corporation sorted the electorate into voter types and predicted their responses. It advised the 1960 Kennedy campaign, and many of its predictions could not be tested.

Paul Green and Vithala Rao brought conjoint measurement into marketing. Respondents rank product profiles, and the method estimates how much each feature is worth to them.

Joshua Epstein and Robert Axtell produced inequality, migration and trade in simulated societies from local rules alone.

The Santa Fe Institute described Swarm, one of the first general-purpose software libraries for building agent-based simulations.

Donella Meadows ranked twelve places to intervene in a complex system, from changing numbers such as taxes to changing the goals and assumptions behind the system.

AnyLogic combined system dynamics, agent-based and discrete-event models in one commercial tool.

Facebook likes alone predicted people’s personality traits and their party.

Joon Sung Park and colleagues used a language model to generate thousands of community members and their posts, so designers could see how a social platform might be used before launch.

Twenty-five language-model agents lived in a simulated town, kept memories, planned their days and organised a party between them.

Language models prompted with demographic backstories reproduced the opinion distributions of US subgroups, and predicted individuals’ votes from their backstories. The backstories held demographics only, not each person’s own answers.

The opinions language models give by default differed from those of many US demographic groups.

Agents grounded in an interview with each person reproduced 85% of that person’s own consistency when retested two weeks later.

Synthetic survey answers varied less than real ones, and many regression results estimated on them changed size or sign.

Twin-2K-500 scored twins at 0.717 ordinal alignment. A mega-study on the same people found that on new stimuli twins correlated with their humans at r ≈ 0.20.

Fine-tuning on large sets of human responses brought models closer to the spread of real answers, though closer matching can lower single-answer accuracy. One model was trained to predict behaviour across many experiments.

Models prompted with an identity reflected what outsiders think of a group more than what its members say, and treated each group as uniform.

Twins built from German panel microdata reached 0.73 ordinal alignment on average and 0.788 at best. Anacreon, a mixture of models each fine-tuned on similar people, reported a median of 0.775 on one customer audience.